Practical Statistics Formulas: Key Equations and When to Use Them
A concise guide to common statistics formulas, when to use them, and worked checks you can run. Covers mean, variance, z-scores, t-tests, regression basics.

Quick answer: core statistics formulas and their use
Statistics formulas provide compact tools for describing data (mean, median), measuring spread (variance, standard deviation), standardizing values (z-scores), testing hypotheses (t-tests, chi-squared), and modeling relationships (linear regression coefficients). For everyday analysis, start with descriptive formulas to summarize your sample, then use inferential formulas when you need to generalize from a sample to a population.
A practical path: compute the mean and standard deviation to understand center and spread, use z-scores or t-statistics to compare observations or sample means, and apply regression formulas to estimate a linear relationship. Which exact formula you choose depends on whether the data are paired or independent, the sample size, and whether you can assume a normal distribution.
- Mean (x̄): arithmetic average; quick center-of-mass for numeric data.
- Median: middle value; robust when outliers skew the mean.
- Variance (s^2) and Standard Deviation (s): measure spread around the mean.
- Z-score: (x - μ) / σ; standardizes a value versus a known population.
- T-statistic: (x̄ - μ0) / (s/√n); for small samples with unknown population σ.
- Confidence interval (CI): range for a parameter; CI width depends on sample size and variance.
- Linear regression slope (β̂): covariance(x,y) / variance(x); predicts change in y per unit of x.
What these statistics formulas actually tell you
Each formula compresses an idea about the data into a single number or set of numbers. The mean summarizes central tendency assuming numeric values are on a comparable scale; if the distribution is heavily skewed the mean can be misleading, and the median or trimmed mean may better represent a 'typical' observation. Variance and standard deviation quantify how tightly values cluster around the mean: a large standard deviation means wide spread, small means values are tightly grouped.
Standardizing with z-scores or t-statistics converts raw measurements into units of standard deviation. That makes different variables or samples directly comparable: a z-score of 2 means the value is two population-standard-deviations above the mean. T-statistics use the sample standard deviation when the population standard deviation is unknown, and their sampling distribution is wider when sample sizes are small.
Confidence intervals and hypothesis tests use the same underlying formulas but answer different questions: confidence intervals estimate a plausible range for a parameter (e.g., a population mean), while hypothesis tests weigh evidence against a specific claim (e.g., mean = μ0). Regression formulas estimate how one variable changes with another and provide standard errors that let you test whether a slope is distinguishable from zero.
Key clues
Before plugging numbers into a formula, inspect the raw data and context. Confirm the measurement scale (continuous, ordinal, binary), check for obvious data-entry errors or outliers, and note whether observations are independent. Many formulas assume independent observations; paired designs (before/after, matched samples) require paired versions of t-tests or other adjustments.
Look at sample size and distribution shape. Small samples (n < 30 as a rough rule) make normality assumptions and variance estimates more consequential. If the distribution is visibly non-normal—heavy tails, skew, or multimodality—consider robust alternatives (median-based measures, bootstrapped intervals) or transformations before using formulas that rely on normality.
- Scale: Are values continuous, counts, proportions, or categories? Use formulas appropriate to the scale.
- Independence: Are observations independent? Paired or clustered data need different formulas.
- Sample size: Small n increases uncertainty and changes which distributions to use (t vs z).
- Outliers: Extreme values can skew means and inflate variance; check with boxplots or trimmed summaries.
- Missing data: Listwise deletion changes sample size and estimates—consider imputation when missingness is nonrandom.
- Units and measurement error: Confirm units (minutes vs seconds) and consider measurement precision in variance estimates.
- Contextual plausibility: Does a calculated CI or effect size make sense given the sampling method and domain knowledge?
Step-by-step workflow for using statistics formulas
1) Define your question and match it to an estimation or test problem. Are you estimating a mean, comparing two groups, testing association, or predicting an outcome? That decision narrows the candidate formulas (mean/CI, two-sample t-test, chi-squared, regression).
2) Inspect and prepare the data: check for typos, convert units, flag missing values, and visualize distributions. Simple plots—histograms, boxplots, scatterplots—highlight issues that change which formulas are appropriate.
3) Choose the estimator or test and its assumptions. For a single mean with unknown σ and a moderate sample, use the t-based CI: x̄ ± t* (s/√n). For a proportion, use the Wilson or Agresti–Coull interval rather than the naive normal approximation when n or p is small.
4) Compute the formula with transparent bookkeeping: record sample size, degrees of freedom, test statistic, and p-value or interval endpoints. For regression, report slope, intercept, standard errors, R², and residual diagnostics to justify model adequacy.
5) Interpret results in context: translate test statistics into practical terms (effect sizes, predicted values) and check sensitivity to assumptions (e.g., do results change with outlier removal or a nonparametric method?).
6) Report uncertainty clearly. A 95% CI does not guarantee the parameter lies inside for any single sample; it expresses long-run coverage under repeated sampling with the same procedures.
- Start: define parameter (mean, proportion, difference, slope).
- Prepare: clean data, visualize, check independence and scale.
- Select: choose formula and distribution (z vs t, chi-squared, F).
- Compute: show intermediate values—x̄, s, n, degrees of freedom, t/z scores.
- Diagnose: examine residuals or bootstrap estimates if assumptions worry you.
- Report: give numbers, uncertainty, and practical interpretation.
Examples: worked statistics formulas for common situations
Example 1 — Sample mean and 95% CI: Suppose you measured the time students took on a test: n = 25, sample mean x̄ = 48.4 minutes, sample standard deviation s = 6.2. The 95% t-critical value with df = 24 is about 2.064. The CI is x̄ ± t*(s/√n) = 48.4 ± 2.064*(6.2/5) = 48.4 ± 2.56, so (45.84, 50.96) minutes. This interval communicates plausible population means given the sample and variance.
Example 2 — Two-sample t-test (unequal variances): Group A (n1 = 20, x̄1 = 75, s1 = 8), Group B (n2 = 22, x̄2 = 70, s2 = 10). The difference in means is 5. The standard error for the difference uses both variances: SE = sqrt(s1^2/n1 + s2^2/n2) ≈ sqrt(64/20 + 100/22) ≈ sqrt(3.2 + 4.55) ≈ 2.82. The t-statistic is 5 / 2.82 ≈ 1.77. Compare to the appropriate degrees-of-freedom approximation (Welch’s df) to get a p-value and judge evidence strength.
Example 3 — Z-score standardization: If the population mean μ = 100 and σ = 15, an observed score x = 130 has z = (130−100)/15 = 2.0, meaning it lies two population standard deviations above average. Use z-scores to combine or compare measurements that have different units or scales.
Example 4 — Simple linear regression slope: With paired observations (xi, yi), the slope β̂ = Σ(xi−x̄)(yi−ȳ) / Σ(xi−x̄)^2. If the covariance is 14.8 and the variance of x is 7.4, the slope is 2.0: for each one-unit increase in x, y increases on average by two units. Always check residuals for nonlinearity and heteroskedasticity before trusting predictions.
- CI formula (t): x̄ ± t* (s/√n) — use when σ unknown and sample moderately small.
- Two-sample t (Welch): t = (x̄1−x̄2) / sqrt(s1^2/n1 + s2^2/n2) — handles unequal variances.
- Z-score: z = (x−μ)/σ — standardizes observations to population σ.
- Regression slope: β̂ = cov(x,y) / var(x) — needs linearity and independent residuals.
Limitations: where formulas are fragile and how to verify
Formulas are only as reliable as their inputs and assumptions. Normal-theory CIs and t-tests assume approximate normality of sampling distributions; with small samples or extreme skew, those assumptions fail and p-values and intervals can be misleading. In such cases consider nonparametric methods (Wilcoxon tests, bootstrap intervals) or transformations that stabilize variance.
Independence violations—clustered, hierarchical, or time-series data—invalidate simple variance formulas. Standard errors computed as if observations were independent will be too small, inflating false-positive rates. Use cluster-robust standard errors, mixed models, or time-series techniques when data structure requires them.
Measurement error and biased sampling affect point estimates and intervals. If the sampling design oversamples particular groups or the measurement instrument adds noise, report that limitation and, where possible, apply design-based corrections or sensitivity analyses. When decisions are consequential, verify results with replication, larger samples, or subject-matter experts.
- Small samples: switch to t-based methods or bootstrap; avoid normal approximations for proportions when n or p is small.
- Non-independence: check design and use appropriate clustered or mixed-model formulas.
- Outliers and heavy tails: robust estimators or trimmed means reduce sensitivity.
- Biased samples and measurement error: disclose and, if possible, adjust using known design weights or instrument calibration.
Related guides
Next step: try a worked calculator or check the solver guide
If you want a hands-on check of these formulas, try the Statistics Calculator at https://statisticsai.app/tools/statistics-calculator for mean, SD, CI, t-tests, and regression steps. For insight into automated solution checks and how to validate results, see the guide on how the solver works: https://statisticsai.app/guides/ai-statistics-problem-solver. Use these tools as a first pass—always confirm with your data diagnostics and domain knowledge before making decisions.
Take the next step →Frequently asked questions
When should I use a z-score versus a t-statistic?
Use a z-score when the population mean and population standard deviation are known and the sampling distribution can be treated as normal (often large samples). In practice, the population standard deviation is rarely known; when σ is unknown and estimated by the sample s, and especially when sample size is moderate or small, use a t-statistic with degrees of freedom equal to n−1 for a single mean or the appropriate Welch df for two-sample comparisons.
How do I calculate variance and why square deviations?
Variance is the average squared deviation from the mean: for a sample, s^2 = Σ(xi−x̄)^2 / (n−1). Squaring deviations prevents positive and negative differences from canceling and gives more weight to larger deviations. The standard deviation (the square root of variance) brings the measure back to the original units, which is often easier to interpret.
What formula should I use for proportions and their confidence intervals?
For a sample proportion p̂ = x/n, the classic approximate CI uses p̂ ± z*√(p̂(1−p̂)/n). When n is small or p̂ near 0 or 1, the normal approximation performs poorly; prefer Wilson or Agresti–Coull intervals or use exact (Clopper–Pearson) methods for conservative coverage. These alternatives reduce misleading intervals in edge cases.
Can I rely on formulas for causal claims from observational data?
No—formulas compute associations and quantify sampling uncertainty, but causal claims require assumptions about confounding, selection, and temporal order that math alone cannot verify. Use careful study design (randomization, natural experiments) or causal inference methods (adjustment, matching, instrumental variables) and consult domain experts before interpreting associations as causal effects.
